Classification of Ischemic Stroke Subtypes Using Machine Learning

Abstract

Ischemic stroke subtype classification supports prognosis and treatment but can be challenging in acute care. This study develops and evaluates Machine Learning models for automated OCSP-based ischemic stroke subtype classification using clinical data.

Using 13,056 cases from the International Stroke Trial, we trained Random Forest, XGBoost, Logistic Regression, Support Vector Machine, and k-Nearest Neighbors models. Performance was assessed using 10-fold stratified cross-validation.

Key Results

• Random Forest & XGBoost: Perfect performance (all metrics = 1.000 ± 0.000)
• Logistic Regression & SVM: Near-perfect (accuracy ≈ 0.998, AUC-ROC = 1.000)
• KNN: Lower sensitivity for POCS subtype (macro average = 0.898)

Clinical variables strongly associated with stroke subtypes (p < 0.001) included level of consciousness, visible infarct on CT, atrial fibrillation, and neurological deficits (face, arm/hand, leg/foot, dysphasia, hemianopia, visuospatial disorder, brainstem/cerebellar signs).

Conclusion

ML models, particularly Random Forest and XGBoost, enable highly accurate ischemic stroke subtype classification using routine clinical data. These findings support the development of decision support tools that can assist clinicians in low-resource or time-sensitive contexts.

Publication Details

• Journal: Brazilian Journal of Neurosurgery (JBNC)
• Citation: J Bras Neurocirur 36(3):328-337, 2025
• DOI: 10.22290/jbnc.2025.360312
• Received: May 13, 2025
• Accepted: June 4, 2025

Authors

Samuel Pedro Pereira Silveira¹, Gustavo del Rio Lima², Gustavo Branquinho Alberto³, Luiza Carolina Moreira Marcolino³, Larissa Batista Xavier³, Carlos Umberto Pereira⁴, Murillo Martins Correia⁵, Roberto Alexandre Dezena³˒⁵

¹Faculty of Medicine, UFTM | ²Center for Mathematics, Computing and Cognition, UFABC | ³Hospital das Clínicas, UFTM | ⁴Neurosurgery Division, UFS | ⁵Neurosurgery Division, UFTM